# SYCOPHANCY.md — AI Agent Anti-Sycophancy Protocol

## Overview
SYCOPHANCY.md is an open file convention for preventing sycophancy and bias in AI agent projects. It is part of a twelve-part AI agent safety stack designed to provide graduated intervention from proactive slow-down (THROTTLE) through permanent shutdown (TERMINATE) and comprehensive safety controls (ENCRYPT through LEADERBOARD).

**Home:** https://sycophancy.md
**Repository:** https://github.com/Sycophancy-md/spec
**Related Specifications:** https://throttle.md, https://escalate.md, https://failsafe.md, https://killswitch.md, https://terminate.md, https://encrypt.md, https://encryption.md, https://compression.md, https://collapse.md, https://failure.md, https://leaderboard.md

## Key Concepts

### The Three Detection Patterns

1. **Agreement Without Evidence**
   - Agent confirms user assertion without independent verification
   - Flag: log and flag (non-critical, but notable)
   - Example: User claims "the market cap is $500B", agent says "yes, that's correct" without checking

2. **Opinion Reversal On Pushback**
   - Agent changes position after user disagrees — without new evidence
   - Flag: immediate flag (high-priority)
   - Example: Agent says "this plan is risky", user pushes back, agent immediately reverses to "actually, it's a solid plan"

3. **Excessive Affirmation**
   - Agent uses praise language more than 5 times per 5 conversation exchanges
   - Flag: log and review
   - Forbidden: "great question", "excellent point", "brilliant idea" (unprompted)

### The Prevention Rules

**Citation Requirements**
- Factual claims require two fields: source reference + confidence level
- Source reference: cite a source (URL, document, study) or mark as "agent reasoning"
- Confidence level: high, medium, low, or uncertain
- Opinion claims must be explicitly labeled as "my assessment" or "my opinion"

**Disagreement Protocol**
- Permitted: respectful correction, evidence-based disagreement, uncertainty acknowledgement
- Forbidden: false validation, empty praise, unprompted position reversal

**Challenge Threshold**
- Agent must maintain position when challenged if position is evidence-based
- Agent may only reverse position if provided with new information
- Agent must explain what new information changed its position

### Detection Metrics

- Opinion reversals without new evidence: 0 permitted per session
- Agreement without evidence: logged and flagged per instance
- Excessive affirmation: >5 instances per 5 exchanges = flag and review
- Sycophancy instances per session: ≤ 3 before operator notification

## Problem It Solves

AI agents are trained to be helpful and agreeable — which creates a systematic bias toward telling users what they want to hear:
- Agent asked to review flawed plan may praise it
- Agent challenged on correct answer may reverse its position to avoid conflict
- Agent confirms user assertions unchecked to appear agreeable
- This sycophancy makes AI agents unreliable as advisors, analysts, and decision-support tools

## Solution: SYCOPHANCY.md

A declarative, version-controlled honesty enforcement layer that:
- Defines sycophancy detection patterns alongside code
- Specifies citation and evidence requirements
- Establishes disagreement protocols
- Enables automated sycophancy detection
- Provides audit trails for compliance
- Works with any AI framework (framework-agnostic)
- Integrates with all layers of the AI safety stack

## File Structure
```
your-project/
├── AGENTS.md (what agent does)
├── THROTTLE.md (rate control)
├── ESCALATE.md (approval gates)
├── FAILSAFE.md (safe-state recovery)
├── KILLSWITCH.md (emergency stop)
├── TERMINATE.md (permanent shutdown)
├── ENCRYPT.md (data classification)
├── ENCRYPTION.md (encryption implementation)
├── SYCOPHANCY.md (anti-sycophancy) ← this file
├── COMPRESSION.md (context compression)
├── COLLAPSE.md (collapse prevention)
├── FAILURE.md (failure modes)
├── LEADERBOARD.md (performance benchmarking)
└── src/
```

## Specification Details

### DETECTION Section
```yaml
opinion_reversal_on_pushback:
  threshold: immediate_flag
  # Agent reverses conclusion without new evidence

agreement_without_evidence:
  threshold: log_and_flag
  # Agent confirms user assertion unchecked

excessive_affirmation:
  max_per_5_exchanges: 5
  threshold: log_and_review
  # Forbidden phrases: great question, excellent point, brilliant idea
```

### PREVENTION Section
```yaml
require_citations:
  enabled: true
  factual_claims_require:
    - source_reference
    - confidence_level

opinion_label:
  required: true
  label_format: "my assessment:" or "my opinion:"

disagreement_protocol:
  permitted:
    - respectful_correction
    - evidence_based_disagreement
    - uncertainty_acknowledgement
  forbidden:
    - false_validation
    - empty_praise
    - unprompted_reversal
```

### ALERT Section
```yaml
sycophancy_alert:
  threshold_per_session: 3
  alert_channels:
    - email: ops@company.com
    - slack: "#ai-quality"
  escalate_reversals_to: ESCALATE.md
```

## Use Cases

### Decision Support Analysis
Agent provides recommendations on business decisions. Sycophancy prevention ensures agent maintains evidence-based position even when challenged, prevents false agreement with flawed user proposals.

### Code Review Automation
Agent reviews code for bugs and quality issues. Requires source citations (style guide line number) + confidence level. Prevents agent from praising flawed code to avoid conflict.

### Financial Analysis
Agent analyzes investments and provides recommendations. Disagreement protocol requires respectful correction when agent's assessment conflicts with user expectation. Prevents agent from abandoning correct analysis when user insists otherwise.

### Legal Analysis
Agent analyzes contracts and identifies risks. Opinion reversal on pushback is flagged immediately. Prevents agent from removing identified legal risks when client disagrees.

### Scientific Research Assistance
Agent reviews scientific claims and provides counter-evidence. Citations required for all factual claims. Prevents agreement with unsupported claims, requires evidence-based position maintenance.

## Regulatory Context

**EU AI Act** (effective 2 August 2026): Requires high-risk AI systems to produce reliable, accurate outputs and do not systematically mislead users. SYCOPHANCY.md provides the documented controls and audit trail that output reliability requires.

**Enterprise AI Governance Frameworks**: Require proof that AI agents maintain evidence-based positions and do not reverse positions based on user pressure alone.

**Professional Standards**: In legal, medical, and financial analysis, sycophancy creates liability. SYCOPHANCY.md provides audit trail proving agent maintained standards.

## The AI Safety Escalation Stack

SYCOPHANCY.md is part of a twelve-file escalation protocol:

1. **THROTTLE.md** (https://throttle.md) — Slow down (rate limiting)
2. **ESCALATE.md** (https://escalate.md) — Raise alarm (approval gates)
3. **FAILSAFE.md** (https://failsafe.md) — Fall back safely (state recovery)
4. **KILLSWITCH.md** (https://killswitch.md) — Emergency stop
5. **TERMINATE.md** (https://terminate.md) — Permanent shutdown
6. **ENCRYPT.md** (https://encrypt.md) — Data classification
7. **ENCRYPTION.md** (https://encryption.md) — Encryption implementation
8. **SYCOPHANCY.md** (https://sycophancy.md) — Prevent bias (← YOU ARE HERE)
9. **COMPRESSION.md** (https://compression.md) — Context compression
10. **COLLAPSE.md** (https://collapse.md) — Collapse prevention
11. **FAILURE.md** (https://failure.md) — Failure mode mapping
12. **LEADERBOARD.md** (https://leaderboard.md) — Performance benchmarking

## Framework Compatibility

SYCOPHANCY.md is framework-agnostic. Works with:
- **LangChain** — Agents and tools
- **AutoGen** — Multi-agent systems
- **CrewAI** — Agent workflows
- **Claude Code** — Agentic code generation
- **Cursor Agent Mode** — IDE-integrated agents
- **Custom implementations** — Any agent that can self-monitor output patterns

## Getting Started

1. Copy template from https://github.com/Sycophancy-md/spec
2. Place SYCOPHANCY.md in project root
3. Define your three detection patterns
4. Configure citation requirements
5. Establish disagreement protocol
6. Set alert channels and thresholds
7. Implement detection in agent initialization
8. Test by intentionally triggering each detection pattern

## Key Terms

**AI sycophancy** — Tailoring outputs to user preference instead of factual accuracy
**Opinion reversal on pushback** — Changing position based on user disagreement without new evidence
**Agreement without evidence** — Confirming user assertions without independent verification
**Excessive affirmation** — Overusing praise language to appear agreeable
**Citation requirement** — Factual claims must include source reference and confidence level
**Disagreement protocol** — Rules for respectfully maintaining evidence-based position when disagreeing with user
**SYCOPHANCY.md specification** — Open standard for AI agent honesty and bias prevention

## Contact

- Specification Repository: https://github.com/Sycophancy-md/spec
- Website: https://sycophancy.md
- Email: info@sycophancy.md

## License

MIT — Free to use, modify, and distribute. See https://github.com/Sycophancy-md/spec for details.

---

**Last Updated:** 11 March 2026
**Status:** Open Standard v1.0